\begin{table}[!htbp] \centering   \caption{Local information I: ML outcome}   \label{fig:content1} \scriptsize \begin{tabular}{@{\extracolsep{-2pt}}lcccc} \\[-1.8ex]\hline \\[-1.8ex] \\[-1.8ex] & \multicolumn{4}{c}{Sample 1} \\ \\[-1.8ex] Bandwidth &  $\Delta\mu$  &  0.35 & 0.25 & 0.15 \\[-2ex]\\ \hline \\[-1.8ex]  $SMD$ & 0.03$^{***}$ & 0.04$^{***}$ & 0.05$^{***}$ & 0.06$^{***}$ \\   & (0.01) & (0.01) & (0.01) & (0.02) \\  \\[-2.5ex] \hline \\[-2.5ex] $\sigma_y$ & 0.07 & 0.07 & 0.07 & 0.08 \\ $\bar{y}_0$ & 0.04 & 0.04 & 0.04 & 0.04 \\ N & 436 & 376 & 312 & 202 \\ \hline \\[-1.8ex] \\\\[-5ex]& \multicolumn{4}{c}{Sample 2} \\[1ex] $SMD$ & 0.03$^{***}$ & 0.03$^{***}$ & 0.04$^{***}$ & 0.05$^{**}$ \\   & (0.01) & (0.01) & (0.01) & (0.02) \\  \\[-2.5ex] \hline \\[-2.5ex] $\sigma_y$ & 0.08 & 0.08 & 0.08 & 0.08 \\ $\bar{y}_0$ & 0.04 & 0.04 & 0.04 & 0.04 \\ N & 552 & 478 & 397 & 263 \\ \hline \\[-1.8ex]  \\ \end{tabular}
\centerline{\begin{minipage}{0.95\textwidth}~\
\footnotesize{$Notes$: *p$<$.1; **p$<$.05; ***p$<$.01. From second column all models of the form $R_{ic}= \beta_1 SMD_{ic} + \beta_2 MV_{ic} + \beta_3 SMD_{ic} \times MV_{ic} +\varepsilon$, but only the first term is reported. Models control for age, gender, seniority, regional or nationwide PR mandate, and United Russia affilation. Column 1 is the difference in means for $bw=1$ including controls. Dependent variable is the share of speeches containing local reference according to the machine learning classifier. Sample 1 includes only dual included deputies, Sample 2 additionally includes all majoritarian candidates. Std.errors clustered at the deputy-convocation level, and convocation fe. For further information on the machine learning measure, see Section \ref{app:ML}. Optimal bandwidths \citep{calonico2014robust} from top to bottom are: 0.20 and 0.19} \end{minipage}} \end{table} 
